TODO:¶
- Create widget and deploy
- Get more examples of not-poison-oak
- Get better metric to track than accuracy (really care about maximizing recall, less about precision)
- Clean data more
Imports and set Bing API Key¶
In [86]:
from fastai2.vision.all import *
from fastai2.vision.widgets import *
from utils import *
import shutil
key = 'bfe4c5c0c789435b9894b166b1ef6d6a'
Process Data¶
Load Data¶
In [122]:
path = Path('data'); path.ls()
Out[122]:
In [19]:
def search_images_bing(key, term, count, min_sz=128):
client = api('https://api.cognitive.microsoft.com', auth(key))
return L(client.images.search(query=term, count=count, min_height=min_sz, min_width=min_sz).value)
In [20]:
labels = ['poison-oak', 'leaves']
counts = [150, 450]
for label, count in zip(labels, counts):
os.makedirs(path/label, exist_ok=True)
res = search_images_bing(key, label, count)
download_images(path/label, urls=res.attrgot('content_url'))
In [21]:
(path/labels[0]).ls(), (path/labels[1]).ls()
Out[21]:
Verify images¶
In [124]:
verify_images(get_image_files(path))[-5:]
Out[124]:
In [25]:
bad = verify_images(get_image_files(path))
print(bad)
for img in bad: img.unlink()
In [26]:
(path/'poison-oak').ls(), (path/'leaves').ls()
Out[26]:
Create DataBlock¶
In [125]:
data = DataBlock(blocks=(ImageBlock, CategoryBlock),
get_items=get_image_files,
get_y=parent_label,
splitter=RandomSplitter(valid_pct=0.2, seed=42))
Apply data augmentation¶
In [138]:
aug_data = data.new(item_tfms=RandomResizedCrop(224), batch_tfms=aug_transforms())
Create dataloader¶
In [139]:
dls = aug_data.dataloaders(path, bs=64)
In [144]:
dls.show_batch(max_n=5, nrows=1, unique=True)
Look at batch¶
In [140]:
xb, yb = first(dls.valid)
In [141]:
xb.shape, yb.shape
Out[141]:
In [142]:
dls.show_batch(max_n=9, figsize=(9, 9))
In [143]:
len(dls.train_ds), len(dls.valid_ds)
Out[143]:
Modeling¶
Create Model¶
In [145]:
learn = cnn_learner(dls, resnet34, metrics=accuracy)
Train model¶
In [146]:
learn.fine_tune(4)
In [151]:
learn.fine_tune(4)
Analysis¶
In [152]:
interp = ClassificationInterpretation.from_learner(learn)
Plot results¶
In [153]:
learn.show_results(max_n=9, figsize=(9, 9))
Confusion matrix¶
In [154]:
interp.plot_confusion_matrix()
Top Losses¶
In [155]:
interp.plot_top_losses(5, figsize=(15, 15))
Clean Dataset¶
Image Cleaner widget¶
In [66]:
cleaner = ImageClassifierCleaner(learn, height=256)
In [67]:
cleaner
Make sure to delete/relabel images before moving onto next class / dataset. Image cleaner's tracking resets!
Delete images¶
In [107]:
cleaner.delete()
Out[107]:
In [93]:
PILImage.create(cleaner.fns[3]).show(figsize=(15, 15))
Out[93]:
In [108]:
for i in cleaner.delete(): cleaner.fns[i].unlink()
Relabel images¶
In [119]:
cleaner.change(), cleaner.fns[14]
Out[119]:
In [120]:
for i, new_lbl in cleaner.change(): shutil.move(str(cleaner.fns[i]), path/new_lbl/'relabeled.jpg')
Inference¶
In [158]:
img = PILImage.create('poison-oak.jpg')
img.to_thumb(440)
Out[158]:
In [159]:
learn.predict(img)
Out[159]: